genuinely poor fossil record and this is corroborated by relatively long confidence
intervals on stratigraphic occurrence data in, for example, chondrichthyans, which have a
confidence bracket of >1.8 myr. However, this calculation is based on the compilation of
Ordovician and Silurian occurrences, only 9 of the 67 of which are Ordovician, and these
occurrences are limited to a narrow interval in the Caradoc (Harding Sandstone and its
equivalents). Thus, the fossil record of this group appears to be particularly intermittent
early on, and a reassessment of confidence limits on first appearance based upon the
Ordovician record alone results in a bracket of over six million years. The same situation
is true of thelodonts. Inverse correlation between low SCI (P>0.95) and high GER (P<0.
95), as in the placoderms and actinopterygians, probably results from a good, but poorly
understood, fossil record. The fossil record of placoderms is rich, but attempts to resolve
the relationships of the group have thus far proved only variably successful (e.g. Goujet
and Young 1995; Goujet 2001). The fossil record of actinopterygians is more gap than
record, hence the strongly negative RCI (P<0.99), but the sum of implied ghost ranges is
very low compared with the maximum, and very close to the minimum possible by
optimizing stratigraphic fit/discordance to the tree (GER 0.79; same P as for RCI).
Problems with assessing the quality of the record
There are two potential problems with regard to this analysis, one relating to the analysis
itself, and the second relating to potential artefact in the dataset. First, there is a very poor
correlation, absolutely and proportionally, between the confidence intervals on each of
the groups, which are derived from internal assessments of the quality of the record
within each of the plesions, and the inferred ghost lineages, which are based on analysis at
plesion level (compare Figures 10.4 and 10.5). Paul (1998) suggested that this may be an
appropriate means of identifying ghost lineages that are an artefact of cladistic
methodology, rather than reflecting a true gap in the temporal record of a lineage. We
outlined earlier why we think that our analysis is not subject to this kind of artefact
(plesions are monophyletic).
The second problem relates to the dataset and has implications for the analysis of
confidence intervals and, in turn, their degree of fit to ghost lineages. The calculation of
classic confidence intervals assumes that fossil recovery potential is random. Testing this
assumption is very difficult when dealing with global compilations of palaeontological data
and probably represents the greatest limitation upon the extension of confidence intervals
to global datasets. Nevertheless, there is some evidence to suggest that there are two
significant biases in the dataset, indicating that the existing dataset is not a random sample
of the fossil record. First, the vast majority of known occurrences are from northern
Europe, the USA, and South-East Asia, compared with a global fossil collection bias for
north-west Europe and North America (e.g. Smith 2001). Although there are numerous
fossil records from North America as a whole, the vast majority of taxonomic treatments
of North American faunas (especially Arctic Canada) are new taxa, suggesting that
although the North American record is being recovered rapidly, it has been sampled only
sparsely to date (using the collecting curve analogy we remain on the steep component of
the curve). A bias against collecting central Asian faunas appears to be supported by
records of spot occurrences in terranes such as Tuva (Afanassieva and Janvier 1985). A
208 THE ORIGIN AND EARLY EVOLUTION OF CHORDATES